0
$\begingroup$

First off I will state that I have looked at similar threads (such as this (referenced later) and this) yet I still don't fully understand. My situation is as follows, I am trying to report the results of the following glm:

glm(GotoPB ~ ZF_Pb + ZF_NotPb, data = DF_DL_5, family = binomial(link = "logit"))

In which the variables are:

GotoPB = A factor with two levels: "N" or "Y"

ZF_Pb = A numeric variable

ZF_NotPb = A numeric variable

My summary is as follows: 

Call:
glm(formula = GotoPB ~ ZF_Pb + ZF_NotPb, family = binomial(link = "logit"), 
    data = DF_DL_5)

Deviance Residuals: 
    Min       1Q   Median       3Q      Max  
-2.4285  -1.1333   0.3953   1.1467   1.7764  

Coefficients:
            Estimate Std. Error z value Pr(>|z|)    
(Intercept)  0.07267    0.12283   0.592 0.554124    
ZF_Pb        0.24729    0.04173   5.926  3.1e-09 ***
ZF_NotPb    -0.03548    0.01057  -3.357 0.000788 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for binomial family taken to be 1)

    Null deviance: 840.93  on 608  degrees of freedom
Residual deviance: 778.57  on 606  degrees of freedom
AIC: 784.57

Number of Fisher Scoring iterations: 5

Initially I understood this as: An increase in ZF_Pb increases the chance that GotoPB = Y. But I am a bit confused on the interpretation of the intercept. Referring back to the thread referenced earlier, which states that the intercept is determined alphabetically, does that mean that my intercept is in fact GoToPB-N.

And following on that, would that mean that an increase in ZF_Pb instead equates the opposite; The higher ZF_Pb the larger the chance that GotoPB = N?

I should state that this last conclusion would be in stark contrast to what we would expect, which is why this has left me rather confused. I am worried that I do not understand these results properly. Could someone help me clarify this?

EDIT:

On advice from a friend I used the effects package and the plot(allEffects(model)) function in order to evaluate the effects in a model plot. This results in: This figure

Moreover, based on output from the allEffects function:

 model: GotoPB ~ ZF_Pb + ZF_NotPb

 ZF_Pb effect
ZF_Pb
        0       7.5        15        22        30 
0.4460287 0.8372560 0.9704774 0.9946413 0.9992555 

 ZF_NotPb effect
ZF_NotPb
        0        10        20        30        40 
0.6283205 0.5424455 0.4539719 0.3683128 0.2902256

Leads me to believe that my initial assumption was correct, and that indeed an increase of ZF_Pb coincides with an increased probability of GotoPB=Y.

Improve this question
$\endgroup$
2
  • $\begingroup$ Your anguish reminds me of when I investigated the sanity of relatively statistically uncomplicated stepwise regression output. Perhaps the same related path to a solution. Start with more rudimentary models and see if you are getting anything rational. If not, ask questions about the data. If yes, proceed a next level with an open mind on varying possible paths to incorporate the data. If all fails, can you simulate data with desired interrelationships and find a tool that nearly accurately returns the theoretical parameters? $\endgroup$
    – AJKOER
    Commented May 22, 2020 at 11:45
  • $\begingroup$ Thank you for your reply. I did investigate the data prior to analyses, created some simple but clear illustrations, which seem to suggest the same as my hypothesis: That an increase in ZF_Pb would result in an increased chance that GotoPB = Y. The figures even seem to match the relative estimates of my model. Which is why I am so confused about this output. I am not quite sure if either A) The output really does not match my expectations, or B) If my original interpretation was correct and I need not worry. I have to admit I don't quite follow your last suggestion, could you expand on that? $\endgroup$
    – R. Iersel
    Commented May 22, 2020 at 13:57

0

Reset to default